acceptodds
Under review as a conference paper at ICLR 2027

Beyond Freshness: Closure-Aware Admission for Reliable Agent Memory

Abstract

External memory enables large language model agents to accumulate and reuse information across long-running interactions and multi-step tasks. However, source updates, corrections, and retractions can invalidate derived summaries, plans, and cross-record conclusions. Such memories may violate task constraints even when individual updates are locally correct, records appear fresh, or different maintenance schedules converge to the same state. We study memory-use admission under evolving information: deciding which derived memories remain valid for a request under explicit dependencies and task constraints, while balancing correctness, service availability, and maintenance cost. Direct version checking can miss transitive invalidation; invalidation alone does not restore service; and global rejection and repeated full validation can introduce unnecessary blocking and redundant computation, respectively. We propose Closure-Aware Admission (CAA), which combines task-constraint validation with transitive dependency closure to quarantine inadmissible memories while preserving unaffected service. A persistent incremental implementation reuses validation results that remain valid after updates. Controlled language-model experiments expose the compositional failures motivating this task. Further evaluations across mechanism families, source-driven replay, and native execution on SQLite and Mem0 in raw-storage mode show that CAA matches the decisions of the full transitive-closure baseline under the evaluated contracts, avoiding unsafe admissions missed by direct version checks and unnecessary service losses caused by global rejection. Across the evaluated SQLite and Mem0 workloads, persistent CAA reduces aggregate online execution time relative to persistent full-closure validation while preserving the same admission decisions. These results demonstrate the benefits of reusing valid verification results for efficient and reliable memory maintenance.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large Language Model Agents, Agent Memory, Memory Reliability

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_Rg4ROpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes closure-aware admission (CAA), which combines task-constraint validation with transitive dependency closure to provide efficient and reliable memory maintenance. Evaluations on SQLite and Mem0 show that CAA can reduce aggregate online execution time while preserving the same admission decisions.

Critical strengths
  1. The figure presentations are clear.
  2. Overall results look good.
Critical weaknesses
  1. The biggest concern I have with the submission is the inconsistency between the introduction and the main methodology. While the authors have acknowledged the AI-assisted tools. the gap remains large and has significantly degraded the presentation. For instance, MORPH-MEM is brought up at line 066 without any prior context. Though it seems to be a crucial part of the main methodology, it is not fully explained in the main text, except in Figure 2. This makes me question whether the authors used AI to generate the entire paper without any proper validation.

  2. Another key piece of evidence for this paper being AI-generated is the lack of model appearances and a related work survey. Agent memory usage is in some way highly dependent on the examined models, where large-scale models might behave entirely differently than small models.

  3. The final giveaway is the evaluation section. Throughout the entire section, the paper only brings up the metrics in an abstract way: how many records, how many steps, etc. One of the key characteristics of workloads is input/output lengths. Omitting such important details makes me really question the legitimacy of the paper.

Questions for authors

Please see the weaknesses.

Reviewer_kp9dOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies whether a request may consume maintained memory after source changes. Closure-Aware Admission (CAA) combines declared constraints, failure-localization supports, transitive dependency closure, and request read sets to isolate affected memories. A persistent implementation caches rule outcomes and reuses closure results under a fixed-graph, complete-change-stream contract.

The paper reports agreement with full-closure admission, selective protection on 477 state–request inputs, repository-state replay, and online timing reductions on SQLite and Mem0. The implementation and cost attribution appear coherent within the stated contract. However, the main paper does not establish sufficient methodological novelty beyond conventional dependency maintenance and incremental evaluation. Its application evidence also does not demonstrate a task where maintaining the derived memory provides a net benefit over directly accessing current sources.

This assessment covers the supplied main paper and references. The separately referenced Supplementary Material was not available for inspection.

Critical strengths
  1. A coherent admission and caching contract. Retaining inactive vertices preserves dependency paths. The isolation seed and forward closure are well defined, and the dirty-rule set accounts for uncached, changed-dependent, pending, and unresolved evaluations. I found no apparent inconsistency under the stated deterministic, fixed-graph, complete-change assumptions.

  2. Useful separation of protection and performance. The propagation-depth ablation tests why complete reachability matters, while the rule-recomputation ablation identifies the source of the online savings. Matching storage calls and admission decisions makes the timing comparison more informative.

  3. Transparent unfavorable results. The paper reports that direct source access is cheaper in its repository task, that errors concentrate in three histories, and that closure caching adds only a small benefit beyond rule-result caching.

Critical weaknesses
  1. The algorithmic contribution beyond standard incremental maintenance is insufficiently established

The central constructions are a forward reachability closure, an index from records to affected rules, dirty-result tracking, and reuse of unchanged deterministic evaluations. These are familiar mechanisms in dependency maintenance, build systems, and incremental computation. Prior systems already provide dependency-directed invalidation and selective recomputation; the paper appropriately cites this literature. Adapton, Build Systems à la Carte, Section 4.2.

The experiments do not isolate an additional algorithmic advance. Full transitive quarantine already matches CAA’s protection and availability in Table 2. Section 4.4 then attributes most efficiency improvement to rule-result caching: forcing rule recomputation nearly returns the system to full-recomputation cost. Closure reuse contributes only an additional 0.507% on SQLite and 3.493% on Mem0.

  1. The evaluation does not yet demonstrate the practical value of maintaining derived agent memory

The 477-input suite is useful conformance evidence. Its labels are computed from declared predicates, dependency reachability, and request conditions, and CAA is designed to match the corresponding full-closure policy. Perfect agreement therefore establishes implementation consistency within that contract rather than semantic correctness on independent agent tasks.

The Mem0 integration is also explicitly limited: it uses infer=False, nonsemantic hash embeddings, and exact returned identifiers. This demonstrates storage integration, but does not exercise Mem0’s semantic extraction or retrieval pipeline.

The source-driven application is the most consequential limitation. Direct source access already achieves 182/182 correct services with 356 storage calls, whereas budget-two CAA achieves the same correctness with 728 calls (Table 3; Sections 4.3–4.4). The task is described as copy-based, so maintaining the memo–plan chain introduces stale intermediates without demonstrating a compensating computational or reasoning benefit.

  1. The selective-availability claim requires an important qualification

Equation (5) returns UNKNOWN whenever any rule is unresolved, even if the request reads only a disconnected, validated component.

For example, consider independent components A and B. A request reads only A, its required declarations are complete, and no violation affects A. If a rule confined to B has unresolved bindings, the request still receives UNKNOWN.

This is an intentional conservative policy, not an implementation error. Nevertheless, it limits the claim that CAA preserves unaffected service. Selective isolation applies to localized known violations, while unresolved obligations can still cause global withholding.

Questions for authors

What does CAA add beyond conventional transitive invalidation plus dirty-rule caching under the same declarations and request contract?

Is there an evaluated task in which derived memory performs useful computation and CAA provides a net benefit over direct source access?

How many otherwise unaffected requests receive UNKNOWN because an unrelated rule is unresolved? What scope restriction makes the unaffected-service claim accurate?

Can the authors report workload-level paired timings and repeat dispersion, together with a conventional incremental-validation comparator?

Reviewer_EvmsOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies whether agents should reuse derived memories after their sources change. MORPH-Mem diagnoses maintenance failures by comparing legal operation orders and checking task constraints. Closure-Aware Admission (CAA) evaluates declared rules, propagates inactive records and localized violations through a dependency graph, and checks whether a request intersects the affected region.

Critical strengths

A useful separation of freshness, joint validity, and service availability for agent evaluation.

Critical weaknesses
  1. The technical advance beyond established dependency maintenance and incremental computation is insufficiently demonstrated.

  2. The evaluation validates supplied contracts more convincingly than practical agent-memory reliability.

Questions for authors
  1. What specific technical capability distinguishes CAA from dependency-based truth maintenance?

  2. How are dependencies, failure supports, and the completeness flag obtained and checked in the evaluated applications?

  3. Is there an evaluated workload in which semantic derivatives are necessary or sufficiently expensive that CAA provides an end-to-end benefit over current-source access or on-demand reconstruction? Please identify results that include the cost of constructing the contract, validating it, and repairing affected memories.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.